The Development of Attention Detection Model from Child Behavior for Robot-Assisted Autism Therapy

Using a robot as a mediator in therapeutic activity for autism therapy is a trending solution to reduce the burden for therapists. However, to promote continuous therapy, the use of robots could be extended for assisting parents or caregivers as well. Our group developed a robot called BLISS, which is a mobile robot with friendly toy-like appearance and sensors, to interact with children. The BLISS robot has been used to assist in learning interventions for autistic children through games. Previous studies showed that children paid attention and responded to the robot differently. Inexperienced parents often struggle to provide therapy to their autistic child. If the robot could learn to detect child attention and interact with the child appropriately, could the robot better help maintain child attention to the activity in therapy session? In this study, we developed an attention detection model based on child behavior during activity with the robot. We collected the child behavior data and the child psychologist evaluation for the behavior. We built the attention detection model for 2 levels of attention (high and low) by using k-nearest neighbor classifier, support vector machine, decision tree, and random forest. From the experiment, the random forest model returned the highest accuracy at 70%. We employed the model in the adaptive interaction system where the robot selects action based on child attention. In the pilot experiment, all four participants could stay engaged with the activity for equal to or more than the standard attention span. The promising results showed the potential of using robots to assist parents at home.